EP1886303B1 - Verfahren zum anpassen eines neuronalen netzwerks einer automatischen spracherkennungseinrichtung - Google Patents
Verfahren zum anpassen eines neuronalen netzwerks einer automatischen spracherkennungseinrichtung Download PDFInfo
- Publication number
- EP1886303B1 EP1886303B1 EP05747980A EP05747980A EP1886303B1 EP 1886303 B1 EP1886303 B1 EP 1886303B1 EP 05747980 A EP05747980 A EP 05747980A EP 05747980 A EP05747980 A EP 05747980A EP 1886303 B1 EP1886303 B1 EP 1886303B1
- Authority
- EP
- European Patent Office
- Prior art keywords
- stage
- lhn
- neural network
- phoneme
- linear stage
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 112
- 238000000034 method Methods 0.000 title claims abstract description 35
- 230000006978 adaptation Effects 0.000 claims abstract description 89
- 238000012549 training Methods 0.000 claims abstract description 42
- 239000011159 matrix material Substances 0.000 claims description 36
- 238000004590 computer program Methods 0.000 claims description 6
- 102100024061 Integrator complex subunit 1 Human genes 0.000 claims 2
- 101710092857 Integrator complex subunit 1 Proteins 0.000 claims 2
- 102100033265 Integrator complex subunit 2 Human genes 0.000 claims 2
- 108050002021 Integrator complex subunit 2 Proteins 0.000 claims 2
- 210000002569 neuron Anatomy 0.000 description 56
- 230000006870 function Effects 0.000 description 27
- 238000012360 testing method Methods 0.000 description 26
- 230000003595 spectral effect Effects 0.000 description 16
- 230000004913 activation Effects 0.000 description 12
- 102100026630 Aurora kinase C Human genes 0.000 description 7
- 108090000805 Aurora kinase C Proteins 0.000 description 7
- 238000001228 spectrum Methods 0.000 description 6
- 238000012937 correction Methods 0.000 description 3
- 230000001419 dependent effect Effects 0.000 description 3
- 241000408659 Darpa Species 0.000 description 2
- 230000001427 coherent effect Effects 0.000 description 2
- 230000000052 comparative effect Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 239000013598 vector Substances 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000003780 insertion Methods 0.000 description 1
- 230000037431 insertion Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- NGVDGCNFYWLIFO-UHFFFAOYSA-N pyridoxal 5'-phosphate Chemical compound CC1=NC=C(COP(O)(O)=O)C(C=O)=C1O NGVDGCNFYWLIFO-UHFFFAOYSA-N 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/065—Adaptation
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/16—Speech classification or search using artificial neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
- G10L2015/025—Phonemes, fenemes or fenones being the recognition units
Definitions
- the present invention relates to the field of automatic speech recognition. More particularly, the present invention relates to a method of adapting a neural network of an automatic speech recognition device, a corresponding adapted neural network and a corresponding automatic speech recognition device.
- An automatic speech recognition device is an apparatus which is able to recognise voice signals such as words or sentences uttered in a predefined language.
- An automatic speech recognition device may be employed for instance in devices for converting voice signals into written text or for detecting a keyword allowing a user to access a service. Further, an automatic speech recognition device may be employed in telephone systems supporting particular services, such as providing a user with the telephone number of a given telephone subscriber.
- an automatic speech recognition device performs steps, which will be briefly described herein after.
- the automatic speech recognition device receives the voice signal to be recognised through a phonic channel.
- phonic channels are a channel of a fixed telephone network, of a mobile telephone network, or the microphone of a computer.
- the voice signal is firstly converted into a digital signal.
- the digital signal is periodically sampled with a certain sampling period, typically of a few milliseconds.
- Each sample is commonly termed "frame”.
- successively, each frame is associated to a set of spectral parameters describing the voice spectrum of the frame.
- the pattern matching block calculates the probability that the frame associated to the set of spectral parameters corresponds to that phoneme.
- a phoneme is the smallest portion of a voice signal such that, replacing a first phoneme with a second phoneme in a voice signal in a certain language, two different signifiers of the language may be obtained.
- a voice signal comprises a sequence of phonemes and transitions between successive phonemes.
- phoneme will comprise both phonemes as defined above and transitions between successive phonemes.
- the pattern matching block calculates a high probability for the phoneme corresponding to an input frame, a low probability for phonemes with voice spectrum similar to the voice spectrum of the input frame, and a zero probability for phonemes with a voice spectrum different from the voice spectrum of the input frame.
- frames corresponding to the same phoneme may be associated to different sets of spectral parameters. This is due to the fact that the voice spectrum of a phoneme depends on different factors, such as the characteristics of the phonic channel, of the speaker and of the noise affecting the voice signal.
- Phoneme probabilities associated to successive frames are employed, together with other language data (such, for instance, vocabulary, grammar rules, and/or syntax rules) to reconstruct words or sentences corresponding to the sequence of frames.
- other language data such as, for instance, vocabulary, grammar rules, and/or syntax rules
- the step of calculating phoneme probabilities of an input frame is performed by a pattern matching block.
- the pattern matching block may be implemented through a neural network.
- a neural network is a network comprising at least one computation unit, which is called "neuron”.
- a neuron is a computation unit adapted to compute an output value as a function of a plurality of input values (also called "pattern").
- the neuron transforms the linear sum in [1] according to an activation function g(.).
- the activation function may be of different types. For instance, it may be either a Heaviside function (threshold function), or a sigmoid function.
- This type of sigmoid function is an increasing, [0;1]-limited function; thus, it is adapted to represent a probability function.
- a neural network employed in an automatic speech recognition device is a multi-layer neural network.
- a multi-layer neural network comprises a plurality of neurons, which are grouped in two or more cascaded stages. Typically, neurons of a same stage have the same activation function.
- a multi-layer neural network typically comprises an input stage, comprising a buffer for storing an input pattern.
- an input pattern comprises a set of spectral parameters of an input frame, and sets of spectral parameters of a few frames preceding and following the input frame.
- a pattern typically comprises sets of spectral parameters of seven or nine consecutive frames.
- the input stage is typically connected to an intermediate (or "hidden") stage, comprising a plurality of neurons.
- Each input connection of each intermediate stage neuron is adapted to receive from the input stage a respective spectral parameter.
- Each intermediate stage neuron computes a respective output value according to formulas [1] and [2].
- the intermediate stage is typically connected to an output stage, also comprising a plurality of neurons.
- Each output stage neuron has a number of input connections which is equal to the number of intermediate stage neurons.
- Each input connection of each output stage neuron is connected to a respective intermediate stage neuron.
- Each output stage neuron computes a respective output value as a function of the intermediate stage output values.
- each output stage neuron is associated to a respective phoneme.
- the number of output stage neurons is equal to the number of phonemes.
- the output value computed by each output stage neuron is the probability that the frame associated to the input pattern corresponds to the phoneme associated to the output stage neuron.
- a multi-layer network with a single intermediate stage has been described above.
- a multi-layer network may comprise a higher number of cascaded intermediate stages (typically two or three) between the input stage and the output stage.
- a neural network acquires the ability of computing, for each input frame, the phoneme probabilities, a training of the neural network is required.
- Training is typically performed through a training set, i.e. a set of sentences that, once uttered, comprise all the phonemes of the language.
- Such sentences are usually uttered by different speakers, so that the network is trained in recognizing voice signals uttered with different voice tones, accents, or the like.
- different phonic channels are usually employed, such as different fixed or mobile telephones, or the like.
- the sentences are uttered in different environments (car, street, train, or the like), so that the neural network is trained in recognising voice signals affected by different types of noise.
- a network through such a training set results in a "generalist” neural network, i.e. a neural network whose performance, expressed as a word (or phoneme) recognition percentage, is substantially homogeneous and independent from the speaker, the phonic channel, the environment, or the like.
- an "adapted" neural network may be desirable, i.e. a neural network whose performance is improved when recognising a predefined set of voice signals.
- a neural network may be:
- adaptation set will refer to a predetermined set of voice signals for which a neural network is adapted.
- An adaptation set comprises voice signals with common features, such as voice signals uttered by a certain speaker, as well as voice signals comprising a certain set of words, as well as voice signals affected by a certain noise type, or the like.
- LIN Linear Input Network
- a further adaptation technique presented in this paper is the Retrained Speaker-Independent (RSI) adaptation, wherein, starting from a SI system, the full connectionist component is adapted to the new speaker.
- RSI Retrained Speaker-Independent
- PPN Parallel Hidden Network
- PNN Parallel Hidden Network
- speaker adaptation weights connecting to/from these units are adapted while keeping all other parameters fixed.
- GAMMA a GAMMA approach, wherein the speaker-dependent input vectors are mapped to the SI system (as in the LIN technique) using a gamma filter.
- the Applicant has noticed that the performance of an adapted neural network can be improved over the performance of the neural networks adapted according to the above cited known methods.
- the object of the present invention as claimed in claim 1 is providing a method of adapting a neural network of an automatic speech recognition device allowing to obtain an adapted neural network with improved performance, for a given adaptation set.
- the present invention provides a method of adapting a multi-layer neural network of an automatic speech recognition device, the method comprising the steps of: providing a neural network comprising an input stage for storing at least one voice signal sample, an intermediate stage having input connections associated to a first weight matrix and an output stage having input connections associated to a second weight matrix, said output stage outputting phoneme probabilities; providing a linear stage in said neural network after said intermediate stage, said linear stage having the same number of nodes as said intermediate stage; and training said linear stage by means of an adaptation set, said first weight matrix and said second weight matrix being kept fixed during said training.
- the method of the present invention allows to obtain an adapted neural network with improved performance over a neural network adapted according to the prior art, in particular according to the above cited LIN technique.
- Adaptation according to the present invention is more effective, thus resulting in an increased word/phoneme recognition percentage.
- the step of training said linear stage comprises training the linear stage so that the phoneme probability of a phoneme belonging to voice signals not comprised in said adaptation set is equal to the phoneme probability of said phoneme calculated by said neural network before the step of providing a linear stage.
- a conservative adaptation training advantageously allows to prevent a neural network adapted according to the present invention from loosing its ability in recognising phonemes absent from the adaptation set.
- the adapted neural networks exhibit good performance also in recognising voice signals which are not fully comprised into the adaptation set.
- the further linear stage training is carried out by means of an Error Back-propagation algorithm.
- an equivalent stage could be provided, such an equivalent stage being obtained by combining the further linear stage and either the following intermediate stage or the output stage.
- the present invention as claimed in claim 9 provides a multilayer neural network comprising an input stage for storing at least one voice signal sample, an intermediate stage having input connections associated to a first weight matrix, an output stage having input connections associated to a second weight matrix, and a linear stage which is adapted to be trained by means of an adaptation set, the first weight matrix and the second weight matrix being kept fixed while the linear stage is trained, said output stage being adapted to output phoneme probabilities, wherein said linear stage is provided after said intermediate stage, said linear stage having the same number of nodes as said intermediate stage.
- the present invention as claimed in claim 17 provides an automatic speech recognition device comprising a pattern matching block comprising a neural network as set forth above.
- the present invention as claimed in claim 18 provides a computer program comprising computer program code means adapted to perform all the steps of the above method when the program is run on a computer.
- the present invention as claimed in claim 19 provides a computer readable medium having a program recorded thereon, the computer readable medium comprising computer program code means adapted to perform all the steps of the above method when the program is run on a computer.
- FIG. 1 schematically shows an automatic speech recognition device ASR.
- the automatic speech recognition device ASR comprises a cascade of a front-end block FE, a pattern matching block PM and a decoder DEC.
- the decoder DEC is further connected to a database G, comprising vocabulary, grammar rules and/or syntax rules of the language for which the device ASR is intended.
- the automatic speech recognition device ASR receives from a phonic channel PC a voice signal VS.
- the front-end block FE digitalizes and samples the voice signal VS, thus generating a sequence of frames, and it associates to each frame a respective set of n spectral parameters SP1, ... SPi, ... SPn.
- the spectral parameters SP1, ... SPi, ... SPn are sent to the pattern matching block PM, which in turn outputs phoneme probabilities p(f1), ... p(fk), ... p(fC).
- the phonemes probabilities are sent to the decoder DEC which, according to the information stored into the database G, recognizes the voice signal.
- the pattern matching block PM may comprise a multi-layer neural network.
- Figure 2 schematically shows a three-stage multi-layer neural network.
- the neural network NN of Figure 2 comprises an input stage InS, an intermediate (hidden) stage IntS and an output stage OutS.
- the input stage InS comprises a buffer B, which is adapted to store the pattern SP1, ... SPi, ... SPD of an input frame, which comprises, as already mentioned above, the set of spectral parameters SP1, ... SPi, ... SPn associated to the input frame and sets of spectral parameters associated to a number of frames preceding and following the input frame.
- the intermediate stage IntS comprises a number M of neurons IN1, ... INj, ... INM..
- Each input connection of each neuron IN1, ... INj, ... INM is adapted to receive a respective spectral parameter of the pattern SP1, ... SPi, ... SPD.
- each input connection of each neuron IN1, ... INj, ... INM is associated to a respective weight.
- w ji refers to the weight of the i-th input connection of the j-th intermediate stage neuron. For simplicity, as already mentioned, it is assumed that the bias is zero.
- the output stage OutS comprises a number C of neurons ON1, ... ONk, ... ONC, wherein C is the number of phonemes.
- Each neuron ON1, ... ONk, ... ONC has M input connections.
- Each of the M input connections of each neuron ON1, ... ONk, ... ONC is connected to a respective intermediate stage neuron IN1, ... INj, ... INM.
- each input connection of each neuron ON1, ... ONk, ... ONC is associated to a respective weight.
- w' kj refers to the weight of the j-th input connection of the k-th output stage neuron. Also in this case, for simplicity, it is assumed that the bias is zero.
- the output value computed by each output stage neuron ON1, ... ONk, ... ONC is the probability p(f1), ...p(fk), ... p(fC) according to which the frame associated to the pattern SP1, ... SPi, ... SPD corresponds respectively to the phoneme f1, ... fk, ... fC.
- FIG 3 shows a simplified representation of the three-stage neural network NN of Figure 2 .
- the three stages of the network are represented as rectangles, each rectangle corresponding to a respective stage (InS, IntS, OutS).
- W ⁇ W ⁇ 11 ... W ⁇ 1 ⁇ M ... W ⁇ kj ... W ⁇ CM ... W ⁇ CM .
- Figure 4 shows a known four-stage neural network.
- the neural network of Figure 4 comprises an input stage comprising a buffer (not shown), a first intermediate (hidden) stage IntS1 comprising neurons (not shown), a second intermediate (hidden) layer IntS2 comprising neurons (not shown), and an output stage OutS comprising neurons (not shown).
- the input connections of the first intermediate stage neurons are associated to a weight matrix W.
- the input connections of the second intermediate stage neurons are associated to a weight matrix W'.
- the input connections of the output stage neurons are associated to a weight matrix W".
- Figure 5 shows the three-stage neural network of Figure 3 , adapted according to the present invention.
- the present invention provides for inserting an additional linear stage LHN after an intermediate stage of a neural network.
- Such an additional linear stage LHN comprises a plurality of linear neurons, i.e. neurons with linear activation function.
- the input connections of the additional stage LHN are associated to a weight matrix W LHN , as it will be shown in further details herein after.
- the additional linear stage LHN is placed between the intermediate stage IntS and the output stage OutS.
- the spectral parameters SP1, ... SPi, ... SPD are firstly processed by the weight matrix W and the activation function of the intermediate stage IntS.
- the additional stage LHN performs a linear transform by means of the weight matrix W LHN and the linear activation function.
- the output values estimated by the additional stage LHN are processed by the weight matrix W and the activation function of the output stage OutS, thus resulting in the phoneme probabilities p(f1), ... p(fk), ... p(fC).
- the linear transform performed by the additional linear stage LHN is performed not on the input spectral coefficients, but on the spectral coefficient processed by the intermediate stage. This advantageously increases the impact of the linear transform on the overall neural network operation, thus allowing to obtain an adapted neural network with improved performance.
- the additional stage LHN has a number of neurons which is equal to the number of intermediate stage neurons (M).
- the weight matrix W LHN associated to the input connections of the additional linear stage neurons is optimised by performing an adaptation training by means of an adaptation set. During such an adaptation training, the weight matrixes W and W' are kept fixed.
- the adaptation training is performed through a so-called Error Back-Propagation algorithm as disclosed, for instance, in C. M. Bishop “Neural networks for pattern recognition", Oxford University Press, 1995, pages 140-148 .
- Such an Error Back-Propagation algorithm consists in computing an error function as the difference between the set of computed phoneme probabilities and a set of target phoneme probabilities.
- Such an error function is "back-propagated" through the neural network, in order to compute correction values to be applied to the weights of the weight matrixes. According to the present invention, such correction values are applied only to the weights of the weight matrix W LHN .
- the weight matrix W LHN is a square MxM matrix.
- Both figures 8 and 9 show the four-stage neural network of Figure 4 , which is adapted according to the present invention.
- the additional linear stage LHN is inserted between the first intermediate stage IntS1 and the second intermediate stage IntS2.
- the additional linear stage LHN is inserted between the second intermediate stage IntS2 and the output stage OutS.
- the Applicant has verified that the adapted neural network of Figure 7 has better performance in comparison with the adapted neural network of Figure 6 , as in the network of Figure 7 the additional linear stage LHN performs a linear transform on data which has already been subjected to a greater number of processing operations.
- the weights w LHN pq of the weight matrix W LHN are optimised by performing an adaptation training by means of an adaptation set. During such an adaptation training, the weight matrixes W, W' and W" are kept fixed.
- the adaptation training is performed through an Error Back-propagation algorithm, as described above with reference to Figure 5 .
- the adaptation training of a neural network induces a neural network to compute always a phoneme probability equal to zero for the absent class phonemes.
- the adapted neural network is not able to perform such a task, as the input connection weights optimised through the adaptation training always induce the network to associate a zero probability to that phoneme.
- M.F. BenZeghiba describes a method for overcoming this problem, by adding some examples of phonemes that did not appear in the adaptation data. However, the Applicant has observed that such a method can be improved.
- the additional linear stage weight matrix W LHN is optimised by performing an adaptation training which allows to preserve the performance of the adapted neural network in recognising absent class phonemes.
- the target phoneme probabilities are chosen as follows:
- the absent class phonemes are associated to a target probability which is different from zero, even if it is known a priori that none of the adaptation set frames corresponds to any of these absent class phonemes.
- the target probabilities are preferably chosen so that the target probability of the phoneme corresponding to the frame is substantially higher than the target probability of the absent class phonemes, so that the decoder is induced to consider unlikely that the frame corresponds to an absent class phoneme.
- the weights w LHN pq after the adaptation training are such that the adapted neural network still has the capability of recognising absent class phonemes.
- the additional linear stage LHN may be "absorbed" with the successive stage. More in particular, after computing the optimum weights w LHN pq through an adaptation training, the additional linear stage LHN and the successive stage are optionally replaced by a single equivalent stage.
- the additional linear stage LHN and the output stage OutS may be replaced by a single equivalent stage.
- the Applicant has performed a number of comparative tests between a generalist neural network (i.e. before adaptation), the generalist neural network adapted according to the known LIN technique, and the generalist network adapted according to two different embodiments of the present invention.
- the generalist neural network has been adapted by inserting an additional linear stage (LHN).
- the generalist neural network has been adapted by inserting an additional linear stage which has been trained through conservative adaptation training (LHN+CT)
- the generalist neural network was a four-layer neural network of the type shown in Figure 4 .
- the buffer B size was 273.
- the first intermediate stage comprised 315 neurons, whose activation function g(a) is the sigmoid function defined by equation [2].
- the second intermediate stage comprised 300 neurons, whose activation function g(a) is the sigmoid function defined by equation [2].
- the output stage comprised 683 neurons (for Italian language), whose activation function g(a) is a so-called softmax function, which is a sigmoid function ensuring that the sum of the phoneme probabilities is equal to 1.
- the generalist neural network has been adapted using different adaptation sets, such as:
- Each adaptation set is associated to a respective test set.
- the ensemble of a training set and its respective test set is usually termed "corpus".
- the WSJ0 corpus which has been defined by DARPA Spoken Language Program, has a vocabulary comprising 5000-20000 English words.
- a 5000 word vocabulary has been used.
- a Sennheiser HMD414 microphone has been used, both during the adaptation training and during the tests.
- the WSJ1 Spoke-3 corpus which has been defined by DARPA Spoken Language Program, has a vocabulary comprising 5000 English words.
- the test set comprised 40x8 320 test sentences, uttered by the same ten different non-native speakers.
- the Aurora3 corpus which has been defined by European Union funded SpeechDat-Car project, has a vocabulary comprising 2200 Italian connected digit utterances, divided into training utterances and test utterances. These utterances are affected by different noise types inside a car (high speed good road, low speed rough road, car stopped with motor running, and town traffic).
- the adaptation set used by the Applicant comprised 2951 connected digits utterances, while the test set comprised 1309 connected digits utterances.
- the Consi-12 corpus which has been defined by the Applicant, has a vocabulary comprising 9325 Italian town names.
- the adaptation set used by the Applicant comprised 53713 adaptation utterances, while the test set comprised 3917 test utterances.
- the AppWord corpus which has been defined by the Applicant, has a vocabulary comprising applicative Italian words such as "avanti”, “indietro”, “fine”, or the like.
- the adaptation set used by the Applicant comprised 6189 adaptation utterances, while the test set comprised 3094 test utterances.
- the Digcon corpus which has been defined by the Applicant, is a subset of the SpeechDat corpora.
- the adaptation set used by the Applicant comprised 10998 adaptation utterances, while the test set comprised 1041 test utterances.
- Table 1 reported below shows the results of the tests. Performance is expressed as word recognition percentage. For each adapted network, the performance is evaluated by referring to the test set coherent with the respective adaptation set. For the generalist neural network, performance is evaluated for all the above reported test sets. Table 1 Application Vocabulary Channel Speaker adaptation method Consi-12 Appl. Words Digcon Aurora 3 WSJ0 WSJ1 Spoke-3 none 85.4 96.2 98.6 87.9 82.8 49.7 LIN 88.8 96.6 98.5 94.2 85.2 57.4 LHN 90.4 97.9 99.1 95.0 86.4 70.2 LHN+CT 89.9 97.7 99.0 94.6 87.4 71.6
- Neural network adapted through the known LIN technique has shown improved performance for each adaptation set, except the adaptation set Digcon.
- Performance has been further improved by adapting the generalist network according to the first embodiment of the present invention (LHN).
- LHN generalist network according to the first embodiment of the present invention
- a neural network adapted according to the present invention exhibits better word recognition performance in comparison with neural networks adapted according to the prior art.
- Table 2 shows the results of a further comparative test of Italian continuous speech recognition for some of the above cited adaptation tests. Performance is expressed as speech recognition accuracy, which is obtained by subtracting from the recognised word percentage both the word insertion percentage and the word deletion percentage.
- Table 2 adaptation method Consi-12 App. Words Digcon Aurora3 (4%) (48%) (86%) (86%) none 70.7 LIN 63.7 57.3 23.3 -8.6 LHN 59.4 36.3 -47.3 -52.1 LHN+CT 59.3 54.7 60.6 55.8
- the voice signal comprises both phonemes comprised into the adaptation sets, and absent class phonemes.
- the generalist neural network exhibits a speech recognition accuracy equal to 70.7%.
- the conservative adaptation training advantageously allows to improve the performance.
- the speech recognition accuracy increases from -47.3% (LHN) to 60.6% (LHN-CT), while for the adaptation set Aurora3 the speech recognition accuracy increases from -52.1 % to 55.8%.
- the present invention advantageously allows to obtain, for most of the considered adaptation sets, improved performance in word recognition test performed through test sets coherent with the respective adaptation sets. Besides, an improvement in speech recognition accuracy can be obtained by performing a conservative adaptation training according to a preferred embodiment of the present invention.
Landscapes
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Machine Translation (AREA)
- Computer And Data Communications (AREA)
- Character Discrimination (AREA)
- Feedback Control In General (AREA)
- Telephonic Communication Services (AREA)
Claims (19)
- Verfahren zum Anpassen eines mehrschichtigen neuronalen Netzwerks (NN) einer automatischen Spracherkennungsvorrichtung (ASR), wobei das Verfahren die folgenden Schritte umfasst:- Bereitstellen eines neuronalen Netzwerks (NN), das eine Eingangsstufe (Ins) zum Speichern von mindestens einer Sprachsignalprobe, eine Zwischenstufe (IntS, IntS1, IntS2), die Eingangsverbindungen hat, welche mit einer ersten Gewichtungsmatrix (W) verbunden sind, und eine Ausgangsstufe (OutS) umfasst, die Eingangsverbindungen hat, welche mit einer zweiten Gewichtungsmatrix (W') verbunden sind, wobei die Ausgangsstufe (OutS) Phonemwahrscheinlichkeiten ausgibt;- Bereitstellen einer linearen Stufe (LHN) im neuronalen Netzwerk (NN) nach der Zwischenstufe (IntS, IntS1, IntS2), wobei die lineare Stufe (LHN) dieselbe Zahl von Knoten wie die Zwischenstufe (IntS, IntS1, IntS2) hat; und- Trainieren der linearen Stufe (LHN) mittels eines Anpassungssets, wobei die erste Gewichtungsmatrix (W) und die zweite Gewichtungsmatrix (W') während des Trainings unverändert gelassen werden.
- Verfahren nach Anspruch 1, wobei der Schritt des Trainierens der linearen Stufe (LHN) das Training der linearen Stufe (LHN) derart umfasst, dass die Phonemwahrscheinlichkeit eines Phonems, das zu den Sprachsignalen gehört, die nicht im Anpassungsset enthalten sind, gleich der Phonemwahrscheinlichkeit des Phonems ist, das vom neuronalen Netzwerk (NN) vor dem Schritt des Bereitstellens einer linearen Stufe (LHN) berechnet wurde.
- Verfahren nach Anspruch 2, wobei der Schritt des Trainierens der linearen Stufe (LHN) das Trainieren der linearen Stufe (LHN) derart umfasst, dass die Phonemwahrscheinlichkeit des Phonems, das einer Sprachsignalprobe des Anpassungssets entspricht, durch Subtrahieren der Phonemwahrscheinlichkeiten aller Phoneme, die zu den Sprachsignalen gehören, welche nicht im Anpassungsset von 1 enthalten sind, berechnet ist.
- Verfahren nach Anspruch 3, wobei der Schritt des Trainierens der linearen Stufe (LHN) das Trainieren der linearen Stufe (LHN) derart umfasst, dass die Phonemwahrscheinlichkeit der restlichen Phoneme gleich null gesetzt ist.
- Verfahren nach einem der Ansprüche 1 bis 4, wobei der Schritt des Bereitstellens der linearen Stufe (LHN) den Schritt des Bereitstellens der linearen Stufe (LHN) zwischen der Zwischenstufe (IntS) und der Ausgangsstufe (OutS) umfasst.
- Verfahren nach einem der Ansprüche 1 bis 4, wobei der Schritt des Bereitstellens des neuronalen Netzwerks (NN) den Schritt des Bereitstellens eines neuronalen Netzwerks (NN) umfasst, das zwei Zwischenstufen (Int1, Int2) umfasst, und wobei der Schritt des Bereitstellens der linearen Stufe (LHN) das Bereitstellen der linearen Stufe (LHN) zwischen den zwei Zwischenstufen (IntS1, IntS2) umfasst.
- Verfahren nach einem der vorherigen Ansprüche, wobei der Schritt des Trainierens der linearen Stufe (LHN) den Schritt des Trainierens der linearen Stufe (LHN) mittels eines Fehlerrückführungsalgorithmus umfasst.
- Verfahren nach einem der vorherigen Ansprüche, das ferner einen Schritt des Bereitstellens einer äquivalenten Stufe umfasst, die durch Kombinieren der linearen Stufe (LHN) und entweder der folgenden Zwischenstufe (IntS2) oder der Ausgangsstufe (OutS) erhalten ist.
- Berechnungsmodul für mehrschichtige neuronale Netzwerke (NN), das eine Eingangsstufe (Ins) zum Speichern von mindestens einer Sprachsignalprobe, eine Zwischenstufe (IntS, IntS1, IntS2), die Eingangsverbindungen hat, welche mit einer ersten Gewichtungsmatrix (W) verbunden sind, eine Ausgangsstufe (OutS), die Eingangsverbindungen hat, welche mit einer zweiten Gewichtungsmatrix (W') verbunden sind, und eine lineare Stufe (LHN) umfasst, welche dafür ausgelegt ist, mittels eines Anpassungssets trainiert zu werden, wobei die erste Gewichtungsmatrix (W) und die zweite Gewichtungsmatrix (W') unverändert gelassen werden, während die lineare Stufe (LHN) trainiert wird, wobei die Ausgangsstufe (OutS) dafür ausgelegt ist, Phonemwahrscheinlichkeiten auszugeben, wobei die lineare Stufe (LHN) nach der Zwischenstufe (IntS, IntS1, IntS2) bereitgestellt ist, wobei die lineare Stufe (LHN) dieselbe Zahl von Knoten wie die Zwischenstufe (IntS, IntS1, IntS2) hat.
- Neuronales Netzwerk nach Anspruch 9, wobei die lineare Stufe (LHN) dafür ausgelegt ist, so trainiert zu werden, dass die Phonemwahrscheinlichkeit eines Phonems, das zu Sprachsignalen gehört, die nicht im Anpassungsset enthalten sind, gleich der Phonemwahrscheinlichkeit des Phonems ist, das vom neuronalen Netzwerk (NN) vor dem Bereitstellen einer linearen Stufe (LHN) berechnet wurde.
- Neuronales Netzwerk nach Anspruch 10, wobei die lineare Stufe (LHN) dafür ausgelegt ist, so trainiert zu werden, dass die Phonemwahrscheinlichkeit des Phonems, das einer Sprachsignalprobe des Anpassungssets entspricht, durch Subtrahieren der Phonemwahrscheinlichkeiten aller Phoneme, die zu den Sprachsignalen gehören, welche nicht im Anpassungsset von 1 enthalten sind, berechnet ist.
- Neuronales Netzwerk nach Anspruch 11, wobei die lineare Stufe (LHN) dafür ausgelegt ist, so trainiert zu werden, dass die Phonemwahrscheinlichkeit der restlichen Phoneme gleich null gesetzt ist.
- Neuronales Netzwerk nach einem der Ansprüche 9 oder 12, wobei die lineare Stufe (LHN) zwischen der Zwischenstufe (IntS) und der Ausgangsstufe (OutS) bereitgestellt ist.
- Neuronales Netzwerk nach einem der Ansprüche 9 oder 12, wobei das neuronale Netzwerk (NN) zwei Zwischenstufen (Int1, Int2) umfasst und die lineare Stufe (LHN) zwischen den zwei Zwischenstufen (IntS1, IntS2) bereitgestellt ist.
- Neuronales Netzwerk nach einem der Ansprüche 9 bis 14, wobei die lineare Stufe (LHN) dafür ausgelegt ist, mittels eines Fehlerrückführungsalgorithmus trainiert zu werden.
- Neuronales Netzwerk nach einem der Ansprüche 9 bis 15, wobei das neuronale Netzwerk (NN) eine äquivalente Stufe umfasst, die durch Kombinieren der linearen Stufe (LHN) und entweder der folgenden Zwischenstufe (IntS2) oder der Ausgangsstufe (OutS) erhalten ist.
- Automatische Spracherkennungsvorrichtung (ASR), die einen Mustererkennungsblock (PM) umfasst, der ein neuronales Netzwerk (NN) nach einem der Ansprüche 9 bis 16 umfasst.
- Computerprogramm, das Computerprogrammcodemittel umfasst, welche zum Ausführen aller Schritte jeder der Ansprüche 1 bis 8 ausgelegt sind, wenn das Programm auf einem Computer ausgeführt ist.
- Computerlesbares Medium, das ein Programm hat, welches darauf aufgezeichnet ist, wobei das computerlesbare Medium Computerprogrammcodemittel umfasst, welche zum Ausführen aller Schritte jeder der Ansprüche 1 bis 8 ausgelegt sind, wenn das Programm auf einem Computer ausgeführt ist.
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/EP2005/052510 WO2006128496A1 (en) | 2005-06-01 | 2005-06-01 | Method of adapting a neural network of an automatic speech recognition device |
Publications (2)
Publication Number | Publication Date |
---|---|
EP1886303A1 EP1886303A1 (de) | 2008-02-13 |
EP1886303B1 true EP1886303B1 (de) | 2009-12-23 |
Family
ID=35643212
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP05747980A Active EP1886303B1 (de) | 2005-06-01 | 2005-06-01 | Verfahren zum anpassen eines neuronalen netzwerks einer automatischen spracherkennungseinrichtung |
Country Status (7)
Country | Link |
---|---|
US (1) | US8126710B2 (de) |
EP (1) | EP1886303B1 (de) |
AT (1) | ATE453183T1 (de) |
CA (1) | CA2610269C (de) |
DE (1) | DE602005018552D1 (de) |
ES (1) | ES2339130T3 (de) |
WO (1) | WO2006128496A1 (de) |
Families Citing this family (132)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8677377B2 (en) | 2005-09-08 | 2014-03-18 | Apple Inc. | Method and apparatus for building an intelligent automated assistant |
US7884727B2 (en) | 2007-05-24 | 2011-02-08 | Bao Tran | Wireless occupancy and day-light sensing |
US8249731B2 (en) * | 2007-05-24 | 2012-08-21 | Alexander Bach Tran | Smart air ventilation system |
US10002189B2 (en) | 2007-12-20 | 2018-06-19 | Apple Inc. | Method and apparatus for searching using an active ontology |
US9330720B2 (en) | 2008-01-03 | 2016-05-03 | Apple Inc. | Methods and apparatus for altering audio output signals |
US20100030549A1 (en) | 2008-07-31 | 2010-02-04 | Lee Michael M | Mobile device having human language translation capability with positional feedback |
US8676904B2 (en) | 2008-10-02 | 2014-03-18 | Apple Inc. | Electronic devices with voice command and contextual data processing capabilities |
US10276170B2 (en) | 2010-01-18 | 2019-04-30 | Apple Inc. | Intelligent automated assistant |
US8682667B2 (en) | 2010-02-25 | 2014-03-25 | Apple Inc. | User profiling for selecting user specific voice input processing information |
DE102010043680A1 (de) | 2010-11-10 | 2012-05-10 | Siemens Aktiengesellschaft | Wasserfahrzeug mit einem Querstrahlantrieb |
US9262612B2 (en) | 2011-03-21 | 2016-02-16 | Apple Inc. | Device access using voice authentication |
US10057736B2 (en) | 2011-06-03 | 2018-08-21 | Apple Inc. | Active transport based notifications |
US9235799B2 (en) | 2011-11-26 | 2016-01-12 | Microsoft Technology Licensing, Llc | Discriminative pretraining of deep neural networks |
US8700552B2 (en) | 2011-11-28 | 2014-04-15 | Microsoft Corporation | Exploiting sparseness in training deep neural networks |
US10134385B2 (en) | 2012-03-02 | 2018-11-20 | Apple Inc. | Systems and methods for name pronunciation |
US10417037B2 (en) | 2012-05-15 | 2019-09-17 | Apple Inc. | Systems and methods for integrating third party services with a digital assistant |
US9721563B2 (en) | 2012-06-08 | 2017-08-01 | Apple Inc. | Name recognition system |
US8442821B1 (en) | 2012-07-27 | 2013-05-14 | Google Inc. | Multi-frame prediction for hybrid neural network/hidden Markov models |
US9240184B1 (en) * | 2012-11-15 | 2016-01-19 | Google Inc. | Frame-level combination of deep neural network and gaussian mixture models |
US9477925B2 (en) | 2012-11-20 | 2016-10-25 | Microsoft Technology Licensing, Llc | Deep neural networks training for speech and pattern recognition |
US9230550B2 (en) * | 2013-01-10 | 2016-01-05 | Sensory, Incorporated | Speaker verification and identification using artificial neural network-based sub-phonetic unit discrimination |
DE112014000709B4 (de) | 2013-02-07 | 2021-12-30 | Apple Inc. | Verfahren und vorrichtung zum betrieb eines sprachtriggers für einen digitalen assistenten |
US9454958B2 (en) | 2013-03-07 | 2016-09-27 | Microsoft Technology Licensing, Llc | Exploiting heterogeneous data in deep neural network-based speech recognition systems |
US9466292B1 (en) * | 2013-05-03 | 2016-10-11 | Google Inc. | Online incremental adaptation of deep neural networks using auxiliary Gaussian mixture models in speech recognition |
WO2014197335A1 (en) | 2013-06-08 | 2014-12-11 | Apple Inc. | Interpreting and acting upon commands that involve sharing information with remote devices |
EP3008641A1 (de) | 2013-06-09 | 2016-04-20 | Apple Inc. | Vorrichtung, verfahren und grafische benutzeroberfläche für gesprächspersistenz über zwei oder mehrere instanzen eines digitaler assistenten |
US10176167B2 (en) | 2013-06-09 | 2019-01-08 | Apple Inc. | System and method for inferring user intent from speech inputs |
US20150154002A1 (en) * | 2013-12-04 | 2015-06-04 | Google Inc. | User interface customization based on speaker characteristics |
US10296160B2 (en) | 2013-12-06 | 2019-05-21 | Apple Inc. | Method for extracting salient dialog usage from live data |
US9620108B2 (en) | 2013-12-10 | 2017-04-11 | Google Inc. | Processing acoustic sequences using long short-term memory (LSTM) neural networks that include recurrent projection layers |
US9715875B2 (en) | 2014-05-30 | 2017-07-25 | Apple Inc. | Reducing the need for manual start/end-pointing and trigger phrases |
AU2015266863B2 (en) | 2014-05-30 | 2018-03-15 | Apple Inc. | Multi-command single utterance input method |
US10170123B2 (en) | 2014-05-30 | 2019-01-01 | Apple Inc. | Intelligent assistant for home automation |
US9430463B2 (en) | 2014-05-30 | 2016-08-30 | Apple Inc. | Exemplar-based natural language processing |
US9633004B2 (en) | 2014-05-30 | 2017-04-25 | Apple Inc. | Better resolution when referencing to concepts |
US10373711B2 (en) | 2014-06-04 | 2019-08-06 | Nuance Communications, Inc. | Medical coding system with CDI clarification request notification |
US10754925B2 (en) | 2014-06-04 | 2020-08-25 | Nuance Communications, Inc. | NLU training with user corrections to engine annotations |
US9627532B2 (en) * | 2014-06-18 | 2017-04-18 | Nuance Communications, Inc. | Methods and apparatus for training an artificial neural network for use in speech recognition |
US9818400B2 (en) | 2014-09-11 | 2017-11-14 | Apple Inc. | Method and apparatus for discovering trending terms in speech requests |
US10127911B2 (en) | 2014-09-30 | 2018-11-13 | Apple Inc. | Speaker identification and unsupervised speaker adaptation techniques |
US10074360B2 (en) | 2014-09-30 | 2018-09-11 | Apple Inc. | Providing an indication of the suitability of speech recognition |
US9668121B2 (en) | 2014-09-30 | 2017-05-30 | Apple Inc. | Social reminders |
US9824684B2 (en) | 2014-11-13 | 2017-11-21 | Microsoft Technology Licensing, Llc | Prediction-based sequence recognition |
US10152299B2 (en) | 2015-03-06 | 2018-12-11 | Apple Inc. | Reducing response latency of intelligent automated assistants |
US9886953B2 (en) | 2015-03-08 | 2018-02-06 | Apple Inc. | Virtual assistant activation |
US10567477B2 (en) | 2015-03-08 | 2020-02-18 | Apple Inc. | Virtual assistant continuity |
US9721566B2 (en) | 2015-03-08 | 2017-08-01 | Apple Inc. | Competing devices responding to voice triggers |
US10515301B2 (en) | 2015-04-17 | 2019-12-24 | Microsoft Technology Licensing, Llc | Small-footprint deep neural network |
US10460227B2 (en) | 2015-05-15 | 2019-10-29 | Apple Inc. | Virtual assistant in a communication session |
US11113596B2 (en) | 2015-05-22 | 2021-09-07 | Longsand Limited | Select one of plurality of neural networks |
US10083688B2 (en) | 2015-05-27 | 2018-09-25 | Apple Inc. | Device voice control for selecting a displayed affordance |
US9578173B2 (en) | 2015-06-05 | 2017-02-21 | Apple Inc. | Virtual assistant aided communication with 3rd party service in a communication session |
US11025565B2 (en) | 2015-06-07 | 2021-06-01 | Apple Inc. | Personalized prediction of responses for instant messaging |
US9818409B2 (en) * | 2015-06-19 | 2017-11-14 | Google Inc. | Context-dependent modeling of phonemes |
US20160378747A1 (en) | 2015-06-29 | 2016-12-29 | Apple Inc. | Virtual assistant for media playback |
US10304440B1 (en) * | 2015-07-10 | 2019-05-28 | Amazon Technologies, Inc. | Keyword spotting using multi-task configuration |
CN105139864B (zh) * | 2015-08-17 | 2019-05-07 | 北京眼神智能科技有限公司 | 语音识别方法和装置 |
US10366158B2 (en) * | 2015-09-29 | 2019-07-30 | Apple Inc. | Efficient word encoding for recurrent neural network language models |
US10956666B2 (en) | 2015-11-09 | 2021-03-23 | Apple Inc. | Unconventional virtual assistant interactions |
US10049668B2 (en) | 2015-12-02 | 2018-08-14 | Apple Inc. | Applying neural network language models to weighted finite state transducers for automatic speech recognition |
US10366687B2 (en) * | 2015-12-10 | 2019-07-30 | Nuance Communications, Inc. | System and methods for adapting neural network acoustic models |
US10223066B2 (en) | 2015-12-23 | 2019-03-05 | Apple Inc. | Proactive assistance based on dialog communication between devices |
US10229672B1 (en) | 2015-12-31 | 2019-03-12 | Google Llc | Training acoustic models using connectionist temporal classification |
US11227589B2 (en) | 2016-06-06 | 2022-01-18 | Apple Inc. | Intelligent list reading |
US10049663B2 (en) | 2016-06-08 | 2018-08-14 | Apple, Inc. | Intelligent automated assistant for media exploration |
US10586535B2 (en) | 2016-06-10 | 2020-03-10 | Apple Inc. | Intelligent digital assistant in a multi-tasking environment |
DK179415B1 (en) | 2016-06-11 | 2018-06-14 | Apple Inc | Intelligent device arbitration and control |
DK201670540A1 (en) | 2016-06-11 | 2018-01-08 | Apple Inc | Application integration with a digital assistant |
US10474753B2 (en) | 2016-09-07 | 2019-11-12 | Apple Inc. | Language identification using recurrent neural networks |
WO2018057639A1 (en) | 2016-09-20 | 2018-03-29 | Nuance Communications, Inc. | Method and system for sequencing medical billing codes |
US10043516B2 (en) | 2016-09-23 | 2018-08-07 | Apple Inc. | Intelligent automated assistant |
US11281993B2 (en) | 2016-12-05 | 2022-03-22 | Apple Inc. | Model and ensemble compression for metric learning |
WO2018106971A1 (en) * | 2016-12-07 | 2018-06-14 | Interactive Intelligence Group, Inc. | System and method for neural network based speaker classification |
US11204787B2 (en) | 2017-01-09 | 2021-12-21 | Apple Inc. | Application integration with a digital assistant |
DK201770383A1 (en) | 2017-05-09 | 2018-12-14 | Apple Inc. | USER INTERFACE FOR CORRECTING RECOGNITION ERRORS |
US10417266B2 (en) | 2017-05-09 | 2019-09-17 | Apple Inc. | Context-aware ranking of intelligent response suggestions |
DK201770439A1 (en) | 2017-05-11 | 2018-12-13 | Apple Inc. | Offline personal assistant |
US10726832B2 (en) | 2017-05-11 | 2020-07-28 | Apple Inc. | Maintaining privacy of personal information |
US10395654B2 (en) | 2017-05-11 | 2019-08-27 | Apple Inc. | Text normalization based on a data-driven learning network |
DK179496B1 (en) | 2017-05-12 | 2019-01-15 | Apple Inc. | USER-SPECIFIC Acoustic Models |
DK201770427A1 (en) | 2017-05-12 | 2018-12-20 | Apple Inc. | LOW-LATENCY INTELLIGENT AUTOMATED ASSISTANT |
US11301477B2 (en) | 2017-05-12 | 2022-04-12 | Apple Inc. | Feedback analysis of a digital assistant |
DK179745B1 (en) | 2017-05-12 | 2019-05-01 | Apple Inc. | SYNCHRONIZATION AND TASK DELEGATION OF A DIGITAL ASSISTANT |
DK201770431A1 (en) | 2017-05-15 | 2018-12-20 | Apple Inc. | Optimizing dialogue policy decisions for digital assistants using implicit feedback |
DK201770432A1 (en) | 2017-05-15 | 2018-12-21 | Apple Inc. | Hierarchical belief states for digital assistants |
US10311144B2 (en) | 2017-05-16 | 2019-06-04 | Apple Inc. | Emoji word sense disambiguation |
US10403278B2 (en) | 2017-05-16 | 2019-09-03 | Apple Inc. | Methods and systems for phonetic matching in digital assistant services |
US10303715B2 (en) | 2017-05-16 | 2019-05-28 | Apple Inc. | Intelligent automated assistant for media exploration |
DK179560B1 (en) | 2017-05-16 | 2019-02-18 | Apple Inc. | FAR-FIELD EXTENSION FOR DIGITAL ASSISTANT SERVICES |
US10657328B2 (en) | 2017-06-02 | 2020-05-19 | Apple Inc. | Multi-task recurrent neural network architecture for efficient morphology handling in neural language modeling |
CN110892409B (zh) * | 2017-06-05 | 2023-09-22 | 西门子股份公司 | 用于分析图像的方法和装置 |
US11133091B2 (en) | 2017-07-21 | 2021-09-28 | Nuance Communications, Inc. | Automated analysis system and method |
US10706840B2 (en) | 2017-08-18 | 2020-07-07 | Google Llc | Encoder-decoder models for sequence to sequence mapping |
US10445429B2 (en) | 2017-09-21 | 2019-10-15 | Apple Inc. | Natural language understanding using vocabularies with compressed serialized tries |
US10755051B2 (en) | 2017-09-29 | 2020-08-25 | Apple Inc. | Rule-based natural language processing |
WO2019078885A1 (en) * | 2017-10-20 | 2019-04-25 | Google Llc | PARALLEL EXECUTION OF OPERATIONS OF ACTIVATION UNITS WITH RELEASE |
US11024424B2 (en) | 2017-10-27 | 2021-06-01 | Nuance Communications, Inc. | Computer assisted coding systems and methods |
US10636424B2 (en) | 2017-11-30 | 2020-04-28 | Apple Inc. | Multi-turn canned dialog |
US10733982B2 (en) | 2018-01-08 | 2020-08-04 | Apple Inc. | Multi-directional dialog |
US10733375B2 (en) | 2018-01-31 | 2020-08-04 | Apple Inc. | Knowledge-based framework for improving natural language understanding |
US10789959B2 (en) | 2018-03-02 | 2020-09-29 | Apple Inc. | Training speaker recognition models for digital assistants |
US10592604B2 (en) | 2018-03-12 | 2020-03-17 | Apple Inc. | Inverse text normalization for automatic speech recognition |
US10818288B2 (en) | 2018-03-26 | 2020-10-27 | Apple Inc. | Natural assistant interaction |
US10909331B2 (en) | 2018-03-30 | 2021-02-02 | Apple Inc. | Implicit identification of translation payload with neural machine translation |
US11145294B2 (en) | 2018-05-07 | 2021-10-12 | Apple Inc. | Intelligent automated assistant for delivering content from user experiences |
US10928918B2 (en) | 2018-05-07 | 2021-02-23 | Apple Inc. | Raise to speak |
US10984780B2 (en) | 2018-05-21 | 2021-04-20 | Apple Inc. | Global semantic word embeddings using bi-directional recurrent neural networks |
US10892996B2 (en) | 2018-06-01 | 2021-01-12 | Apple Inc. | Variable latency device coordination |
DK201870355A1 (en) | 2018-06-01 | 2019-12-16 | Apple Inc. | VIRTUAL ASSISTANT OPERATION IN MULTI-DEVICE ENVIRONMENTS |
DK179822B1 (da) | 2018-06-01 | 2019-07-12 | Apple Inc. | Voice interaction at a primary device to access call functionality of a companion device |
US11386266B2 (en) | 2018-06-01 | 2022-07-12 | Apple Inc. | Text correction |
DK180639B1 (en) | 2018-06-01 | 2021-11-04 | Apple Inc | DISABILITY OF ATTENTION-ATTENTIVE VIRTUAL ASSISTANT |
US11076039B2 (en) | 2018-06-03 | 2021-07-27 | Apple Inc. | Accelerated task performance |
US11010561B2 (en) | 2018-09-27 | 2021-05-18 | Apple Inc. | Sentiment prediction from textual data |
US11170166B2 (en) | 2018-09-28 | 2021-11-09 | Apple Inc. | Neural typographical error modeling via generative adversarial networks |
US11462215B2 (en) | 2018-09-28 | 2022-10-04 | Apple Inc. | Multi-modal inputs for voice commands |
US10839159B2 (en) | 2018-09-28 | 2020-11-17 | Apple Inc. | Named entity normalization in a spoken dialog system |
US11475898B2 (en) | 2018-10-26 | 2022-10-18 | Apple Inc. | Low-latency multi-speaker speech recognition |
US11494612B2 (en) * | 2018-10-31 | 2022-11-08 | Sony Interactive Entertainment Inc. | Systems and methods for domain adaptation in neural networks using domain classifier |
US11638059B2 (en) | 2019-01-04 | 2023-04-25 | Apple Inc. | Content playback on multiple devices |
US11348573B2 (en) | 2019-03-18 | 2022-05-31 | Apple Inc. | Multimodality in digital assistant systems |
US11183219B2 (en) * | 2019-05-01 | 2021-11-23 | Sony Interactive Entertainment Inc. | Movies with user defined alternate endings |
DK201970509A1 (en) | 2019-05-06 | 2021-01-15 | Apple Inc | Spoken notifications |
US11423908B2 (en) | 2019-05-06 | 2022-08-23 | Apple Inc. | Interpreting spoken requests |
US11307752B2 (en) | 2019-05-06 | 2022-04-19 | Apple Inc. | User configurable task triggers |
US11475884B2 (en) | 2019-05-06 | 2022-10-18 | Apple Inc. | Reducing digital assistant latency when a language is incorrectly determined |
US11140099B2 (en) | 2019-05-21 | 2021-10-05 | Apple Inc. | Providing message response suggestions |
DK180129B1 (en) | 2019-05-31 | 2020-06-02 | Apple Inc. | USER ACTIVITY SHORTCUT SUGGESTIONS |
US11496600B2 (en) | 2019-05-31 | 2022-11-08 | Apple Inc. | Remote execution of machine-learned models |
US11289073B2 (en) | 2019-05-31 | 2022-03-29 | Apple Inc. | Device text to speech |
US11360641B2 (en) | 2019-06-01 | 2022-06-14 | Apple Inc. | Increasing the relevance of new available information |
US11488406B2 (en) | 2019-09-25 | 2022-11-01 | Apple Inc. | Text detection using global geometry estimators |
Family Cites Families (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5285522A (en) * | 1987-12-03 | 1994-02-08 | The Trustees Of The University Of Pennsylvania | Neural networks for acoustical pattern recognition |
US5168524A (en) * | 1989-08-17 | 1992-12-01 | Eliza Corporation | Speech-recognition circuitry employing nonlinear processing, speech element modeling and phoneme estimation |
JP2964507B2 (ja) * | 1989-12-12 | 1999-10-18 | 松下電器産業株式会社 | Hmm装置 |
JP2979711B2 (ja) * | 1991-04-24 | 1999-11-15 | 日本電気株式会社 | パターン認識方式および標準パターン学習方式 |
WO1994010635A2 (en) * | 1992-11-02 | 1994-05-11 | Boston University | Neural networks with subdivision |
US5794197A (en) * | 1994-01-21 | 1998-08-11 | Micrsoft Corporation | Senone tree representation and evaluation |
JPH08227410A (ja) * | 1994-12-22 | 1996-09-03 | Just Syst Corp | ニューラルネットワークの学習方法、ニューラルネットワークおよびニューラルネットワークを利用した音声認識装置 |
US6324510B1 (en) * | 1998-11-06 | 2001-11-27 | Lernout & Hauspie Speech Products N.V. | Method and apparatus of hierarchically organizing an acoustic model for speech recognition and adaptation of the model to unseen domains |
US6253181B1 (en) * | 1999-01-22 | 2001-06-26 | Matsushita Electric Industrial Co., Ltd. | Speech recognition and teaching apparatus able to rapidly adapt to difficult speech of children and foreign speakers |
GB0028277D0 (en) * | 2000-11-20 | 2001-01-03 | Canon Kk | Speech processing system |
-
2005
- 2005-06-01 EP EP05747980A patent/EP1886303B1/de active Active
- 2005-06-01 WO PCT/EP2005/052510 patent/WO2006128496A1/en active Application Filing
- 2005-06-01 CA CA2610269A patent/CA2610269C/en active Active
- 2005-06-01 ES ES05747980T patent/ES2339130T3/es active Active
- 2005-06-01 DE DE602005018552T patent/DE602005018552D1/de active Active
- 2005-06-01 US US11/921,303 patent/US8126710B2/en not_active Expired - Fee Related
- 2005-06-01 AT AT05747980T patent/ATE453183T1/de not_active IP Right Cessation
Also Published As
Publication number | Publication date |
---|---|
EP1886303A1 (de) | 2008-02-13 |
US8126710B2 (en) | 2012-02-28 |
DE602005018552D1 (de) | 2010-02-04 |
ATE453183T1 (de) | 2010-01-15 |
CA2610269C (en) | 2016-02-02 |
ES2339130T3 (es) | 2010-05-17 |
CA2610269A1 (en) | 2006-12-07 |
US20090216528A1 (en) | 2009-08-27 |
WO2006128496A1 (en) | 2006-12-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
EP1886303B1 (de) | Verfahren zum anpassen eines neuronalen netzwerks einer automatischen spracherkennungseinrichtung | |
O’Shaughnessy | Automatic speech recognition: History, methods and challenges | |
Young | A review of large-vocabulary continuous-speech | |
Loizou et al. | High-performance alphabet recognition | |
US6208964B1 (en) | Method and apparatus for providing unsupervised adaptation of transcriptions | |
Thomas et al. | Cross-lingual and multi-stream posterior features for low resource LVCSR systems. | |
US8090581B2 (en) | Frame erasure concealment technique for a bitstream-based feature extractor | |
EP0700031A1 (de) | Spracherkennung mit Erkennung falscher Wörter | |
US20010051871A1 (en) | Novel approach to speech recognition | |
CN1199488A (zh) | 模式识别 | |
JPH075892A (ja) | 音声認識方法 | |
Revathi et al. | Speaker independent continuous speech and isolated digit recognition using VQ and HMM | |
AU776919B2 (en) | Robust parameters for noisy speech recognition | |
Boite et al. | A new approach towards keyword spotting. | |
Rabiner et al. | Historical Perspective of the Field of ASR/NLU | |
EP0782127A2 (de) | Zeitvariable Merkmalsraumvorverarbeitungsprozedur für Spracherkennung bei Telefonen | |
EP1505572B1 (de) | Spracherkennungsverfahren | |
Artières et al. | Connectionist and conventional models for free-text talker identification tasks | |
JPH10254473A (ja) | 音声変換方法及び音声変換装置 | |
Li | Combination and generation of parallel feature streams for improved speech recognition | |
Sankar et al. | Noise-resistant feature extraction and model training for robust speech recognition | |
Beaufays et al. | Using speech/non-speech detection to bias recognition search on noisy data | |
Dumitru et al. | Vowel, Digit and Continuous Speech Recognition Based on Statistical, Neural and Hybrid Modelling by Using ASRS_RL | |
Deligne et al. | On the use of lattices for the automatic generation of pronunciations | |
Rose et al. | A user-configurable system for voice label recognition |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
17P | Request for examination filed |
Effective date: 20071211 |
|
AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU MC NL PL PT RO SE SI SK TR |
|
RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: GEMELLO, ROBERTO,LOQUENDO S.P.A. Inventor name: MANA, FRANCO,LOQUENDO S.P.A. |
|
RAP1 | Party data changed (applicant data changed or rights of an application transferred) |
Owner name: LOQUENDO S.P.A. |
|
DAX | Request for extension of the european patent (deleted) | ||
17Q | First examination report despatched |
Effective date: 20090202 |
|
GRAP | Despatch of communication of intention to grant a patent |
Free format text: ORIGINAL CODE: EPIDOSNIGR1 |
|
GRAS | Grant fee paid |
Free format text: ORIGINAL CODE: EPIDOSNIGR3 |
|
GRAA | (expected) grant |
Free format text: ORIGINAL CODE: 0009210 |
|
AK | Designated contracting states |
Kind code of ref document: B1 Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IS IT LI LT LU MC NL PL PT RO SE SI SK TR |
|
REG | Reference to a national code |
Ref country code: GB Ref legal event code: FG4D |
|
REG | Reference to a national code |
Ref country code: CH Ref legal event code: EP |
|
REG | Reference to a national code |
Ref country code: IE Ref legal event code: FG4D |
|
REF | Corresponds to: |
Ref document number: 602005018552 Country of ref document: DE Date of ref document: 20100204 Kind code of ref document: P |
|
REG | Reference to a national code |
Ref country code: NL Ref legal event code: VDEP Effective date: 20091223 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: LT Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: SE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: FI Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
REG | Reference to a national code |
Ref country code: ES Ref legal event code: FG2A Ref document number: 2339130 Country of ref document: ES Kind code of ref document: T3 |
|
LTIE | Lt: invalidation of european patent or patent extension |
Effective date: 20091223 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: SI Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: PL Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: AT Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: IS Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20100423 Ref country code: BG Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20100323 Ref country code: NL Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: RO Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: EE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: PT Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20100423 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: CZ Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: SK Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 Ref country code: BE Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: GR Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20100324 Ref country code: CY Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
PLBE | No opposition filed within time limit |
Free format text: ORIGINAL CODE: 0009261 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT |
|
26N | No opposition filed |
Effective date: 20100924 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: MC Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20100630 Ref country code: DK Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
REG | Reference to a national code |
Ref country code: CH Ref legal event code: PL |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: LI Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20100630 Ref country code: CH Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20100630 Ref country code: IE Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20100601 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: HU Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20100624 Ref country code: LU Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20100601 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: TR Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT Effective date: 20091223 |
|
REG | Reference to a national code |
Ref country code: FR Ref legal event code: PLFP Year of fee payment: 12 |
|
REG | Reference to a national code |
Ref country code: FR Ref legal event code: PLFP Year of fee payment: 13 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: IT Payment date: 20170623 Year of fee payment: 13 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: CZ Payment date: 20170425 Year of fee payment: 11 |
|
REG | Reference to a national code |
Ref country code: FR Ref legal event code: PLFP Year of fee payment: 14 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: IT Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20180601 |
|
REG | Reference to a national code |
Ref country code: ES Ref legal event code: FD2A Effective date: 20190913 |
|
PG25 | Lapsed in a contracting state [announced via postgrant information from national office to epo] |
Ref country code: ES Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES Effective date: 20180602 |
|
P01 | Opt-out of the competence of the unified patent court (upc) registered |
Effective date: 20230516 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: FR Payment date: 20230523 Year of fee payment: 19 Ref country code: DE Payment date: 20230523 Year of fee payment: 19 |
|
PGFP | Annual fee paid to national office [announced via postgrant information from national office to epo] |
Ref country code: GB Payment date: 20230523 Year of fee payment: 19 |